How OpenCrane works

Every time an agent does something in OpenCrane — answers you, drafts a document, calls a tool — that work happens as a run: a tracked, disposable unit of execution that OpenCrane admits, watches and records from start to finish. This page is a short tour of what that means in practice, for both kinds of agent.

Personal and managed runs, side by side

	Your personal assistant	A managed agent
Who it acts as	You	Its own service identity
What it can see	Only what's been granted to you	Only what its published revision was configured with
How it starts	You talking to it	A schedule, a trigger, or another authorised caller
Persona	Built through your interview	None — its published configuration is its complete instruction set
Typical use	Drafting, research, day-to-day work for one person	Triaging tickets, nightly reports, org-wide or team-wide jobs

Both kinds go through the same admission and execution machinery below — that's what makes every run auditable the same way, no matter who or what started it.

What happens when work starts

you, or a schedule/trigger
       │
       ▼
OpenCrane checks who's asking and what they're allowed to do
       │
       ▼
OpenCrane freezes exactly what this run may use — before anything executes
       │
       ▼
a fresh, disposable Kubernetes Job carries out the work
       │
       ▼
you see the ordered events, any actions taken, and the final outcome

1. OpenCrane authenticates the caller and resolves the organisation.
2. It checks membership, grants, model access and budget.
3. It freezes the accepted inputs — which tools, skills, knowledge and model this exact run may use — before any container starts. Nothing can widen its own access mid-run.
4. A fresh, bounded Job carries out the work and streams results back.
5. Every tool call OpenCrane executes on the agent's behalf is recorded, and any that needs a human decision pauses for approval.
6. The run reaches a final outcome. The Job that executed it can disappear — the durable record of what ran, what it used and what happened does not.

Why "disposable execution, durable record" matters

The container is a detail; it can crash, get rescheduled, or simply finish and vanish. What you audit, retry or investigate later is the run record — never a Pod name.

The words you'll see

Agent (personal or managed)

The thing that does the work. A personal assistant is yours alone; a managed agent is a shared, narrowly scoped worker your organisation configures. See the distinction.

Run

One tracked execution of an agent — a single conversation turn, a scheduled job, a triggered task. A run keeps its state, how many attempts it's had, exactly what it was allowed to use, and how it ended. Retrying creates another attempt on the same run rather than a disconnected new one.

Skill

A reusable ability you give an agent — drafting a follow-up, reviewing a document, summarising a ticket — published as a versioned, reviewed artifact rather than a loose snippet. → Manage skills

Tool

An action an agent can ask OpenCrane to take in another system. The agent proposes it; OpenCrane authorises, executes and records it. → Manage tools

Organisational knowledge

Facts and documents an agent can recall — personal notes for your assistant, shared knowledge for a managed agent — always filtered through who's allowed to see what. → Connect knowledge

Organisation (silo)

Your company's isolated slice of OpenCrane. Every agent, run, grant and piece of knowledge lives inside your organisation's boundary and never crosses into another customer's. → Organisation boundary

Ready? → Install OpenCrane